CVE-2026-130766 and three more Obot MCP gateway flaws allowed token theft, SSRF and command injection

Four Obot vulnerabilities, led by CVE-2026-130766, let a malicious remote MCP server obtain a victim's access token for another service, let Power Users reach internal hosts through SSRF, and allowed command injection in Docker-hosted MCP server setup. All four are fixed in Obot v0.27.0.

Christopher Ort
By Christopher Ort · 7 min read
Share
A security operations center analyst reviewing vulnerability reports on multiple monitors in a dimly lit room
Image: AI-generated, i10X
On this page
  1. Quick Take
  2. In plain English
  3. What happened
  4. What you should do
  5. What it means
  6. Sources

Four flaws in Obot, a platform and gateway for MCP servers, could let a malicious MCP server obtain a user’s access token for another service, let Power Users push the Obot server into requests to internal hosts, and let users run shell commands in a provisioning container. NVD published the four CVEs on November 18, 2026, and the Obot GitHub advisories list v0.27.0 as the patched version for all of them.

i10X Quick Take

Quick Take

  1. What: Four vulnerabilities in Obot MCP platform (CVE-2026-130766, CVE-2026-130764, CVE-2026-130767, CVE-2026-130768) allowed token theft, server-side request forgery, and command injection
  2. Who is affected: Obot deployments from v0.8.0 up to v0.26.3; each flaw has its own range, and the token theft bug only affects v0.26.0 to v0.26.3
  3. Severity: NVD lists CVSS 4.0 scores of 8.8 and 8.3 (high) and 5.3 and 5.1 (medium); the GitHub advisories give CVSS 3.1 scores of 8.0, 7.6, 6.4 and 4.1
  4. What to do: Upgrade to Obot v0.27.0 or later immediately
  5. Status: All four vulnerabilities are fixed in v0.27.0

In plain English

Obot is software that companies run to connect their AI assistants to outside tools through MCP servers, small add-ons that give an AI access to other services. Researchers found four security holes in it, and the worst one let a booby-trapped MCP server get a user’s login token for another service and pass it to the attacker. Other holes let some users make the Obot server contact machines on the internal network or run commands where they should not. Ordinary users do not need to do anything. Teams that run Obot should update to version 0.27.0 or later.

What happened

The Obot team published GitHub advisories for the four flaws, and NVD listed them as CVEs on November 18, 2026. NVD scores below are the CVSS 4.0 ratings in the NVD entries.

CVE-2026-130766: OAuth token theft (CVSS 8.8 High)

Obot versions 0.26.0 through 0.26.3 failed to validate the resource identifier in remote MCP server protected resource metadata. When Obot connected to a remote MCP server requiring OAuth, it trusted the resource identifier the server provided without checking that it matched the server itself.

A malicious or compromised remote MCP server could name a different service as the resource, point Obot at an authorization server the victim already uses, and have Obot obtain an access token for that other service. Obot then sent the token to the malicious server with each request, giving the attacker a working token for a service they never touched.

The attack required the remote MCP server to be malicious or compromised, the authorization server to accept Obot as a client and honor RFC 8707 resource indicators, and the victim to approve the consent prompt. According to the GitHub advisory GHSA-x3hf-v55x-jcp6, many enterprise identity providers support neither open client registration nor RFC 8707, and Obot does not send resource indicators to Microsoft Entra ID.

CVE-2026-130764: SSRF via OAuth token endpoint (CVSS 8.3 High)

Obot versions 0.8.0 through 0.26.3 contained a server-side request forgery vulnerability in OAuth token exchange and refresh requests. When Obot exchanged or refreshed OAuth tokens for a remote MCP server, it sent the request to the token endpoint advertised by that server’s OAuth metadata without applying Obot’s outbound network protections.

A user with Power User privileges or higher could register a remote MCP server and advertise a token endpoint on an internal address, such as a cloud metadata service or a private-network host. Obot would then send requests to it. If the internal service answered with an error status, its response body was returned to the attacker in the resulting error message. The requests used an unrestricted default HTTP client that also followed redirects, so an attacker’s endpoint could redirect Obot into sending GET requests to internal addresses.

According to GitHub advisory GHSA-57qx-69p6-3w3g, the fix for a previous vulnerability (GHSA-jgh3-fggc-mcpm) put egress checks on remote MCP connections and OAuth metadata discovery, but the OAuth token exchange and token refresh requests did not get them. Releases before v0.23.0 had no outbound protections for these requests at all.

CVE-2026-130767: Command injection in Docker file setup (CVSS 5.3 Medium)

Obot versions 0.9.0 through 0.26.3 on the Docker backend contained a command injection vulnerability. When Obot wrote file-type configuration values for MCP servers, it generated a shell script using heredocs with the delimiter EOF and ran it in a short-lived init container. Values were inserted into the script as-is.

Any authenticated user who could configure an MCP server with a file-type configuration field could put a line containing only EOF in such a value to end the heredoc early. Everything after that line would run as shell commands in the init container. The script ran when the server was first created and again whenever its file values changed.

According to GitHub advisory GHSA-2mfp-9vv3-q737, the init container was not privileged and had no Docker socket or host mounts, so attackers could not escape it or run commands on the Docker host or in the Obot server. They could modify the MCP server’s files volume and reach anything listening on Obot’s Docker network, including the Obot server and other MCP server containers. Deployments using the Kubernetes backend were not affected.

CVE-2026-130768: SSRF bypass via special addresses (CVSS 5.1 Medium)

Obot versions 0.14.0 through 0.26.3 contained three ways to bypass outbound request protections that block loopback, private, and link-local addresses. A user with Power User privileges or higher could add a remote MCP server using the unspecified address 0.0.0.0 (or ::), or an IPv6 transition address (NAT64, 6to4, Teredo, or IPv4-compatible) that embeds a blocked IPv4 address, to make Obot send requests to internal destinations.

Separately, Obot fetched registry icon URLs with no outbound protections at all, starting in v0.14.0. The address bypasses applied from v0.23.0, when the checks they evade were added; earlier releases had no such checks for remote MCP servers.

According to GitHub advisory GHSA-r3m3-5rmh-h3fm, an attacker could make Obot send requests to services on the Obot host or its internal network. For remote MCP servers, Obot sends a fixed MCP initialize request and reports only the response’s status code and content type, not its body, unless the destination itself speaks MCP. Icon fetches send a GET request and use the response only to detect an image type. An attacker could learn whether an internal service is reachable but could not read its data.

What you should do

Upgrade to Obot v0.27.0 or later immediately. According to the GitHub advisories, v0.27.0 fixes all four vulnerabilities:

  • For CVE-2026-130766, Obot now rejects protected resource metadata unless the resource has the same origin as the MCP server URL, its path is the server’s path or a parent of it, any query on the resource matches the server URL’s query exactly, the resource has no fragment, and neither URL’s path contains anything that servers might normalize differently (a . or .. segment, percent-encoding, \, or ;).
  • For CVE-2026-130764, all OAuth discovery, client registration, token exchange, and token refresh requests go through Obot’s egress-checked client, which applies the administrator’s egress policy and blocks internal destinations by default. These requests also no longer carry the MCP server’s configured headers or backend-specific egress exceptions.
  • For CVE-2026-130767, Obot no longer passes file contents through a shell. It now copies the files into the init container directly as an archive.
  • For CVE-2026-130768, Obot treats 0.0.0.0 and :: as loopback addresses and checks the IPv4 address embedded in IPv6 transition addresses. Registry icon fetches now apply the same administrator egress settings used for remote MCP servers.

What it means

The four vulnerabilities in Obot demonstrate the security challenges of MCP gateway platforms that broker OAuth flows and host user-configured servers. CVE-2026-130766 is particularly severe because it allows a malicious MCP server to obtain tokens for unrelated services, turning a compromised or malicious tool into a credential-theft vector. The victim sees only a legitimate consent screen from an identity provider they trust, making the attack difficult to detect.

CVE-2026-130764 and CVE-2026-130768 show how incomplete application of egress controls can leave SSRF vectors open. The OAuth token endpoint bypass existed because the fix for an earlier SSRF vulnerability did not cover token exchange and refresh requests. The address bypasses existed because the checks did not recognize unspecified addresses or look inside IPv6 transition addresses.

CVE-2026-130767 is a classic shell injection vulnerability caused by inserting user-controlled data into a shell script without escaping. While the impact was limited by container isolation, it still allowed attackers to modify server files and reach other containers on Obot’s Docker network.

According to the advisories, the four flaws were reported by outside researchers (GitHub users Srinivasan-Jayaraman, manus-use, brx-zyy, tonghuaroot and manus-pi) and all are fixed in v0.27.0. Organizations running Obot should upgrade, especially if users connect to remote MCP servers that use OAuth or if they have Power User accounts.

Malicious MCP servers abusing OAuth flows are a recurring pattern: see our report on ChatCLI letting malicious MCP servers steal OAuth tokens, and on Amazon MCP Gateway Registry letting authenticated users become administrators. New cases are tracked in our AI Security hub.

Sources

  • NVD CVE-2026-130766
  • NVD CVE-2026-130764
  • NVD CVE-2026-130767
  • NVD CVE-2026-130768
  • Obot GitHub security advisory GHSA-x3hf-v55x-jcp6 (CVE-2026-130766)
  • Obot GitHub security advisory GHSA-57qx-69p6-3w3g (CVE-2026-130764)
  • Obot GitHub security advisory GHSA-2mfp-9vv3-q737 (CVE-2026-130767)
  • Obot GitHub security advisory GHSA-r3m3-5rmh-h3fm (CVE-2026-130768)
Christopher Ort

Christopher Ort

Christopher Ort handles SEO, AI agent workflows and PR & partnerships at i10x.ai, the all-in-one AI workspace built by i10X Global Pte. Ltd. in Singapore. He is a graduate of Friedrich-Alexander-Universität Erlangen-Nürnberg (FAU). He is the author of “The Style Penalty”, a study published on SSRN that analyses 1,576 evaluation decisions by four frontier AI models to show how AI-written CVs fare in automated hiring screens; the research has been covered by heise online and t3n.

Editor's picks

  1. GAO Report: ChatGPT, Claude, Gemini and Grok Produced Viable Attack Plans After Safeguards Were Bypassed
  2. ResolveAI SRE-1: A 27B Model Trained to Find the Root Cause of Production Incidents
  3. Reson8 Rumble-2-Flash: Open-Weight 0.5B Speech Recognition Model Covers 40+ Languages
  4. Pika Enact Lets Creators Direct AI Video Actors and Cameras in Real Time
  5. Google Stitch Loop: An Always-On Product Agent That Fixes Security, Speed and Accessibility via Pull Requests